Ape Out kept the accident

Ape Out, Gabe Cuzzillo, 2019, is hard to picture without the ape, and the ape was not the plan. The plan was a time-looping stealth game about pushing and grabbing along walls. Guards arrived. The grab worked on them. Holding a man as a shield and throwing him was the best thing in the build, so the loops left and a gorilla stayed.

Into the Breach, Subset Games, 2018, is the same story in miniature. A tactics game grew a tile that lit up before the enemy struck, and that light was so much better than the hidden attack that the whole design had to become about it. A game designs itself only in the sense that the build is a better critic than the pitch. You still have to notice which accident is the game, and you still have to delete the pitch. Most accidents are just bugs. The rare one makes the rest of the document look embarrassed.

Following the fun means letting the prototype overrule the plan. You play what you have built, find the part that is genuinely interesting, and rebuild the project around that instead of the pitch.

Ape Out is hard to picture without the ape, and yet the ape was not in the original plan. Gabe Cuzzillo set out to make a time-looping stealth game in which the player pushed and grabbed their way along walls.

Guards arrived, and the push and grab verbs worked on them just as well as on the level geometry. Grabbing a guard, holding him as a hostage, and hurling him into a wall turned out to be the best thing in the build. Cuzzillo followed it. Out went the time loops and the stealth, and the bald man at the center of the game became a 300-pound gorilla on a rampage.

A single good mechanic will often imply the rest of the design, if you let it. One decision in a prototype can force the next five.

Into the Breach began life as a conventional tactics game in the Advance Wars mold, with enemies picking attacks at random and keeping their plans secret until their turn came around. Then one enemy telegraphed its attack by lighting up the tile it was about to hit.

That was the most fun part of the build, so Subset Games rebuilt around it. Telegraphed attacks made everything else follow. If players can see the danger, dodging with your own units is trivial, so the game has to be about defending buildings that cannot move. Defending buildings turns into shoving enemies away from them, which turns into steering enemies into one another.

That chain is what people mean when they say a game partly designs itself. Every choice is still the studio's, but the strongest choices are the ones already implied by the thing on screen.

Ideas rarely arrive whole. They usually start out vague, unremarkable, or lifted from somewhere else, and they only become themselves once you can play them.

Ryan Clark started Crypt of the NecroDancer from a question: what if the split-second pressure of Spelunky ran inside a turn-based dungeon crawler? His prototype gave you about a second to commit to each move. Playing it made the rhythm impossible to miss, and the dungeon crawler turned into a roguelike driven by music.

Rocket League got one of its signature moves the same way, by accident. Building the earlier game, Psyonix wanted a speed boost for its battling cars and implemented it as a physics force applied at the rear of the vehicle. Testers worked out that the same force could throw a car into the air. That was a whole extra dimension of play, so the studio kept it.

Bugs have been promoted to features for as long as games have shipped. Hideki Kamiya spotted one in Onimusha: Warlords that let repeated slashes juggle an enemy in midair. It got patched out of that game, but Kamiya kept working on the idea, and aerial juggling ended up at the heart of Devil May Cry.

None of this works without something playable. The ideas surface while you are coding it, playing it, and breaking it, and the job is to notice them and let them contradict the original plan.

Gunpoint shows the shift plainly. Early on it was about a robot in space dropping fridges on people. Later it was a side-scrolling stealth game with a Deus Ex flavor and a possible hacking minigame about rewiring a building. The moment Tom Francis had the rewiring in a playable state, the real game showed up. Rewiring wanted to be a puzzle game, so the project became one.

The swings are biggest early, while nothing is expensive to throw away. But the habit does not stop being useful once the game has a recognizable shape.

After the core is settled, following the fun steers content, feel, and tone rather than the premise. The designer stops inventing and starts curating.

Jonathan Blow has described the puzzles in Braid as displays for the surprising things his time-travel engine could do. The work is finding the interesting outcomes, tidying them up, and staging them so a player can appreciate them.

SpyParty was reshaped by its own players. Chris Hecker watched people find exploits and play in ways he had not intended, and rather than sanding those edges off, he made mind games and psychological bluffing an official part of the game.

Subnautica found its genre in production. Horror was not the goal at the start, but the ocean, the isolation, the dark, and the things living in it kept pulling the project that way, and the team went where the game was already headed.

The bill for this method is predictability. Searching for the game takes an unknown amount of time, which is why it shows up more often in independent work than in blockbusters run to a fixed calendar.

Tom Francis hit that wall on Heat Signature. Coming off Gunpoint, he expected that a loose premise about boarding spaceships would resolve into a good design on its own. It did, in the end, but only after years spent building ship generation, AI, combat, a galaxy map, an economy, and everything around them. The discovery that combat aboard the ships was the best part came late.

Hence the longer version of the advice: fail fast, then follow the fun. Marc LeBlanc framed it that way to argue for slapping something together quickly and looking at the result, because even a dead end tells you where to go next. A failure is cheap as long as it happens before you have spent months defending the wrong idea.

The practical trick is to make the search cheaper, so the distance between having an idea and playing it is as short as you can make it.

Game jams exist for exactly that pressure. A weekend is long enough to learn whether an idea has any life in it and short enough that dropping it costs nothing. Baba Is You came out of one, where the concept could be built, judged, and binned in a couple of days if it had not worked.

Fast prototyping tools do the same job, and so do paper, LEGO, Dreams, or anything else that lets you test a rule before you polish a surface. When the systems already exist and the open question is content, level editors let discovery scale and let more people throw ideas at the game. The fastest way to settle an argument about a mechanic is usually to get a build running and hand it to someone.

Placeholder art, music, and writing keep attention on the mechanics for the same reason. The first jam build of Don't Starve used Link from Zelda as its main character. Nobody was evaluating the art. The question was whether the survival loop deserved another month.

A search with no boundaries is too big to finish. Fixing one element while everything around it stays negotiable makes the rest of the search tractable.

Journey was built iteratively, but thatgamecompany kept one thing immovable: the game would be about love. That single commitment gave the team a heading and cut down the list of discoveries worth pursuing.

The fixed point can be a theme, a feeling, a constraint, or a player fantasy. It works like a compass. The game is still free to surprise you, and you still have a basis for deciding which surprises are yours to keep.

The takeaway is that waiting for the finished idea is a mistake. Ape Out, Crypt of the NecroDancer, and Crashlands all look so tightly focused that it is natural to assume each began as one clean insight.

They did not. They came from people who started building, played what they made, got things wrong, tripped over bugs, spotted the parts worth keeping, and pulled on those threads. The talent on display is not just having the first idea. It is hearing what the build is telling you, failing quickly, backing the strongest thread, and assembling scattered discoveries into a coherent whole. You are the one playing it, so you are the one who gets to say what is working.

So build something small enough to play. Pay attention to what it does. Follow the fun, and the game may well finish designing itself.

In the Flockbay app, build the plan in one room and play it looking for the action you repeat without being asked. Then delete a system the plan thought was essential and rebuild around the action. If you cannot bear the deletion, the plan might have been right. Play the deletion before you decide.

An AI game maker for game jams and a start-here template keep the room small enough that an accident is visible.